[Mlir-commits] [mlir] [mlir][vector] Drop innermost unit dims on transfer_write. (PR #78554)

[Han-Chung Wang]

================
@@ -1152,8 +1152,71 @@ struct FoldI1Select : public OpRewritePattern<arith::SelectOp> {
   }
 };
 
-// Drop inner most contiguous unit dimensions from transfer_read operand.
-class DropInnerMostUnitDims : public OpRewritePattern<vector::TransferReadOp> {
+/// Returns the number of dims can be folded away from transfer ops. It returns
+/// a failure if strides and offsets can not be resolved.
+static FailureOr<size_t>
+getTransferFoldableInnerUnitDims(MemRefType srcType, VectorType vectorType) {
+  SmallVector<int64_t> srcStrides;
+  int64_t srcOffset;
+  if (failed(getStridesAndOffset(srcType, srcStrides, srcOffset)))
+    return failure();
+
+  // According to vector.transfer_read/write semantics, the vector can be a
+  // slice. It pads the indices with `1` starting from beginning. Thus, we have
+  // to offset the check index with `rankDiff` in `srcStrides` and source dim
+  // sizes.
+  size_t result = 0;
+  int rankDiff = srcType.getRank() - vectorType.getRank();
+  for (int64_t i = 0, e = vectorType.getRank(); i < e; ++i) {
+    // Check that the inner dim size is 1 for both memref/tensor type and
+    // vector slice. It can be folded only if they are 1 and the stride is 1.
+    int dim = vectorType.getRank() - i - 1;
+    if (srcStrides[dim + rankDiff] == 1 &&
+        srcType.getDimSize(dim + rankDiff) == 1 &&
+        vectorType.getDimSize(dim) == 1) {
+      result++;
+    } else {
+      break;
+    }
+  }
+  return result;
+}
+
+/// Returns a MemRef type that drops inner `dimsToDrop` dimensions from
+/// `srcType`.
----------------
hanhanW wrote:

Yes, this is only apply to the trailing dims.

https://github.com/llvm/llvm-project/pull/78554